An Avalon mining on P2Pool with even 3 modules does indeed take its CPU to the limit. The only real difference between P2Pool and other pools is that P2Pool's generation transaction is much larger. I'm guessing that the bottleneck is hashing the generation transaction ~16 times per second (though that isn't much...). If that's true, perhaps cgminer could be optimized to compress everything before the Stratum nonce to a SHA-256 midstate?
There's an obvious optimisation there that I didn't spot and it's only now that we have hardware being outstripped by the work generation requirements that I've looked into it. Thanks for the pointer, I shall see what I can do.
Yes, that did it. I've committed code to the cgminer git tree which dramatically reduces the cpu usage (confirmed on my avalon mining to p2pool). When my server comes back online I'll upload new avalon firmware.